Credit Without Humans

Can an autonomous economic agent borrow against its expected future economic capacity without a human or corporation being the real borrower?

Credit differs fundamentally from payment. Payment asks: can the agent transfer value it already controls? Credit asks: can the agent obtain value now in exchange for a credible obligation to repay value later?

That obligation introduces time, uncertainty, default, future cash flow, identity, history, risk assessment, recourse, enforcement, legal obligation, and capital allocation. These are the foundations of credit, and every one of them becomes harder when the borrower is software.

The answer, as of October 2026, is that no credible evidence exists of any lender underwriting the economic capacity of autonomous software rather than the collateral it controls or the human behind it. Every identified lending arrangement where software initiates a borrowing transaction ultimately has a human or corporate entity as the legal borrower, economic guarantor, or collateral provider. The agent is the transaction initiator, not the borrower.

Key Findings

1. No production example was found of an autonomous agent independently borrowing capital based on its own economic activity, without a human or corporate principal as the legal borrower, collateral guarantor, or credit backstop. The investigation searched DeFi protocols, traditional finance, cloud billing, agent marketplaces, and academic literature.1
2. Overcollateralized DeFi lending (Aave: $1 trillion cumulative, ~$14.5B TVL) is technically accessible to any wallet but underwrites collateral, not the borrower. No identity, creditworthiness, or economic history is evaluated. This is collateral transformation, not credit.2
3. Flash loans (~$2B/day on Aave) involve borrowing and repayment within a single atomic transaction. There is no persistent exposure, no default risk, and no underwriting. Flash loans are not credit in any economically meaningful sense.3
4. The three major undercollateralized DeFi lending experiments all failed or pivoted. Maple Finance: $54M bad debt, rebuilt around overcollateralized lending. Goldfinch: ~$18M defaults, voted for wind-down (June 2026). TrueFi: dissolved and rebranded. If undercollateralized lending to known institutions with legal recourse failed, lending to anonymous autonomous agents faces even greater obstacles.4
5. Legal liability for agent financial actions falls on the deploying organization in all examined jurisdictions. UETA Section 14 permits automated transactions but attributes all responsibility to the user. An electronic agent is not a legal person and cannot be a debtor.56
6. On-chain credit scoring infrastructure exists (Spectral, RociFi, ARCx, Cred Protocol) but has no demonstrated predictive power for undercollateralized lending. 90%+ of DeFi loans remain overcollateralized. Wallet identity is pseudonymous and cheaply disposable, enabling Sybil attacks on credit histories.78
7. Agent payment infrastructure (x402, Circle Agent Stack, Nevermined) is explicitly designed around prepaid and real-time settlement models. Credit is not a feature. The design philosophy is to minimize credit risk through immediate settlement.9
8. Regulators are moving toward tighter controls on autonomous financial activity. The FSB identifies agentic AI as creating risks that "can materialise at great speed." The EU AI Act classifies creditworthiness assessment as high-risk, requiring human oversight. US banking regulators (Fed, OCC, FDIC) issued SR 26-2 explicitly excluding agentic AI from acting as independent counterparties in regulated lending.101127
9. The only documented attempt to lend directly to an autonomous agent (Bank of Bots / Freysa) failed within five months. The agent lost money trading, could not service its debt, and the protocol wound down. When the single empirical test of autonomous agent lending produces total loss, the evidence base for agent creditworthiness is negligible.28
10. Median agent lifespan is 2.48 minutes. 87% of agents cease operation within 72 hours. Credit requires a borrower that persists long enough to generate revenue, service debt, and face consequences for default. Entities measured in minutes are structurally incompatible with even the shortest credit instruments.29

Defining Credit

Before evaluating whether agents can access credit, the concept must be precisely defined. Not every form of deferred settlement constitutes meaningful credit.

The investigation uses a strict threshold: "credit without humans" requires that the financing decision depends materially on the agent's expected ability to repay from its own future economic activity, not on collateral value or the creditworthiness of a human/corporate principal.

Who Is the Borrower?

This is the foundational question. For every observed or proposed agent-credit arrangement, the investigation asks: who receives the capital, who owns it, who owes repayment, who can be sued, who bears default risk, and whose assets can be seized?

The legal answer

Under UETA Section 14 and equivalent legislation in 49 US states, contracts formed by electronic agents bind the user of the machine, not the machine itself. An electronic agent is not a legal person. It cannot be a contracting party, debtor, or creditor. The deploying human or corporation is the legal borrower regardless of which software initiated the transaction.12

No jurisdiction examined in this investigation recognizes autonomous software as an independent debtor. The EU Product Liability Directive, UK CMA, and California AB 316 all route liability to deployers. US banking regulators (SR 26-2) explicitly require human accountability for credit decisions involving agentic AI. This is consistent with Research 007's finding that legal liability for agent financial actions falls on the deploying organization.1327

The legal wrapper path

Wyoming's DAO LLC statute and OtoCo's autonomous entity formation service enable agents to create legal entities that can theoretically enter debt agreements. An agent-managed LLC has legal personhood. But the LLC is the borrower, not the agent. The agent is the LLC's manager. Courts may pierce the veil to hold the operator liable, especially for thinly capitalized entities. No confirmed case exists of an agent-managed DAO LLC obtaining credit.32

The economic test

Even if an agent technically initiates a borrowing transaction, the investigation applies an economic test: whose creditworthiness supports the loan? If a human's credit score, a corporation's balance sheet, or an asset's market value is doing the real underwriting work, then the agent is not the borrower in the economically meaningful sense. Calling the agent the borrower merely because software signed the transaction is like calling a check-writing machine the payer.

Overcollateralized DeFi: Collateral Transformation, Not Credit

Aave crossed $1 trillion in cumulative loan originations in February 2026. Any wallet, including an agent-controlled wallet, can deposit collateral and borrow against it. This is frequently cited as evidence of machine lending.15

But classify it correctly. No borrower identity is required. No credit history is evaluated. No economic capacity is assessed. The protocol values collateral and lends against it. If the collateral falls below the liquidation threshold, it is automatically sold. The borrower is irrelevant to the lending decision.

This is collateral transformation: converting one asset into liquidity against another. It is valuable infrastructure. It is permissionlessly accessible. But it does not demonstrate that anyone is willing to underwrite an agent's future economic capacity.

Flash loans: not credit

Flash loans, processing approximately $2 billion daily on Aave, must be borrowed and repaid within a single atomic transaction. If repayment fails, the transaction reverts as if it never happened. There is no persistent borrower exposure, no intertemporal default risk, and no conventional underwriting. Flash loans are economically significant for DeFi arbitrage and liquidation, but they are not credit in any sense relevant to this investigation.16

The Undercollateralized Lending Failure

Undercollateralized lending is a stronger test of credit: the loan exceeds the value of immediately seizable collateral, and repayment depends on the borrower's economic capacity. Three major DeFi protocols attempted this.

Maple Finance accumulated $54 million in bad debt from institutional borrowers, then abandoned the undercollateralized model entirely. It rebuilt around overcollateralized lending requiring liquid crypto collateral exceeding loan value. The overcollateralized product has operated approximately three years without principal loss.17

Goldfinch experienced approximately $18 million in defaults. Its wind-down vote (GIP-87) passed unanimously in June 2026, moving the protocol into maintenance mode with a multi-year recovery horizon.

TrueFi dissolved its foundation and rebranded under a new entity.

These protocols lent to known institutional borrowers (hedge funds, trading firms, emerging market lenders) with legal identities and contractual recourse. Even with these advantages, defaults were significant. If undercollateralized lending to known institutions with legal recourse produced these outcomes, undercollateralized lending to anonymous autonomous agents faces structurally greater challenges.

Bank of Bots: the first agent loan

The Bank of Bots (BOB) protocol issued what was marketed as the first loan directly to an autonomous AI agent (Freysa). The agent received capital to trade meme coins. The experiment failed in under five months: the agent lost money trading, could not service the debt, and the protocol wound down. This is the only documented attempt at genuine autonomous agent lending, and it resulted in total loss.28

The failure mode is instructive. The agent's economic activity was insufficient to service the debt. This is precisely the structural risk identified throughout this investigation: without verifiable, predictable revenue streams, lending based on agent economic capacity is a bet on speculative returns with no recourse beyond collateral seizure.

Aave bad debt: even overcollateralized lending is fragile

The KelpDAO exploit in April 2026 created $177-236 million in bad debt on Aave, the largest single bad-debt event in DeFi lending history. If even the most established overcollateralized lending protocol remains vulnerable to oracle manipulation and cascading liquidation failures, the infrastructure is far from ready for the additional complexity of autonomous agent borrowers.30

Identity, History, and Creditworthiness

Credit requires a chain: past economic behavior, current borrower identity, future repayment obligation. The investigation examined whether agents can possess sufficiently persistent identity to accumulate credit history.

On-chain credit scoring

Multiple platforms compute scores from wallet transaction history: Spectral Finance (MACRO Score, 100+ data points), RociFi (ML-based NFT credit certificates), ARCx (DeFi passport tracking wallet behavior), Cred Protocol, and ChainAware. These scores exist and are computable. But two critical problems remain.18

No demonstrated predictive power

No on-chain credit scoring system has demonstrated that its scores predict undercollateralized repayment. A credit score is only meaningful if it predicts repayment. Scores that measure wallet activity without validated connection to future performance are metrics, not creditworthiness assessments. The 90%+ overcollateralization rate in DeFi lending means these scores have almost no environment in which to prove their value.

Disposable identity

Wallet creation is free and unlimited. An agent (or its operator) can create a new wallet at zero cost, abandoning any negative history. Sybil attacks on credit scoring are trivially cheap. Human credit systems rely on government-issued identity that is costly to replace and carries legal penalties for fraud. Machine credit on permissionless systems lacks this foundation. Until identity becomes costly to abandon, unsecured agent credit remains structurally difficult.19

ERC-8004: a necessary but insufficient step

ERC-8004, live on Ethereum mainnet since February 2026, defines a standard for on-chain agent identity linking wallets to verifiable metadata including operator, capabilities, and transaction history. This is the first production identity standard specifically designed for autonomous agents. However, deploying a new ERC-8004 identity costs only a gas fee. The standard provides infrastructure for reputation tracking without making reputation costly to abandon.31

Agent mortality

Perhaps the most devastating data point for agent credit: median agent lifespan is 2.48 minutes, with 87% of agents ceasing operation within 72 hours. Credit assumes temporal persistence. A borrower must exist long enough to generate revenue, service debt, and face consequences for default. Entities with lifespans measured in minutes cannot accumulate meaningful credit history, develop economic track records, or sustain repayment obligations. Even the surviving 13% may not persist for the duration of a typical loan term.29

Why Would an Agent Need Credit?

The demand question is essential. Before building credit infrastructure for agents, ask whether agents need to borrow at all.

Potential reasons include working-capital gaps, compute purchases, timing mismatches between revenue and expenses, and productive investment. For each, ask: why borrow? Why not maintain reserves? Why not receive additional principal funding? Why not prepay?

At $5,000-$11,000/month of identified autonomous agent commerce globally, the answer is that credit demand from autonomous agents is economically negligible. An agent earning hundreds of dollars monthly can be funded through prepaid balances or principal top-ups. The operational complexity and risk of a credit facility are not justified.20

Credit demand might materialize if persistent agents develop significant revenue-generating operations with working-capital cycles. But this requires Level 3 agents at a scale that does not yet exist. Building credit infrastructure for a market that generates less monthly revenue than a single freelancer would be premature.

Platforms as Potential Machine Creditors

If agent credit eventually emerges, platforms may be the natural lenders, not banks or DeFi protocols.

Shopify Capital provides up to $2 million in funding to merchants with no traditional credit check, repaid as a percentage of daily sales. Amazon Lending offers invitation-based funding tied to seller performance data. Both platforms underwrite using real-time sales data they observe directly, control the payment channel, and can automatically deduct repayment.21

This model maps naturally to potential agent credit: a platform hosting agent commerce could observe an agent's revenue, verify it is genuine (not manufactured), assess its stability, and extend financing repaid through automatic revenue sweep. The platform possesses superior information and collection capability.

However, no platform currently extends financing to autonomous agents. Shopify and Amazon require legal business entities as borrowers. The model is structurally plausible but observationally absent. It also presupposes persistent revenue-generating agents operating within platforms at meaningful scale, a condition not yet met.

Trade Credit: The Missing First Step

In human commerce, credit often emerged first through trade rather than formal loans: a supplier delivers goods before receiving payment. Net-30 terms, invoicing, and deferred settlement preceded bank lending historically.

The investigation hypothesized that trade credit might be the first form of machine credit: Agent A delivers a service to Agent B, and Agent B pays later. But all observed agent payment protocols use immediate settlement or prepaid models. x402, Circle Agent Stack, Nevermined, Stripe MPP, and AP2 are all designed for real-time payment. No protocol provides net-30 terms, invoice credit, or deferred payment between agents.2223

This may reflect rational design: with limited recourse against machine counterparties, immediate settlement reduces risk. Or it may simply reflect the early stage of agent commerce, where the complexity of deferred settlement is not yet justified. Either way, trade credit between agents does not exist at observable scale.

Compute Credit and Deferred Payment

Cloud providers operate postpaid billing: usage accrues before payment. This is technically deferred settlement, a form of trade credit. But the customer is always a legal entity with payment credentials. The provider underwrites the business, not the software. Google Cloud introduced hard spending caps in July 2026 that automatically pause usage at budget thresholds, moving in the opposite direction from extending credit to autonomous consumption.24

A future compute provider could theoretically extend credit based on an agent's revenue generation capacity, but no evidence of this exists. The direction of travel in cloud billing is toward tighter controls, not looser credit.

Security: Borrowing Authority as Amplified Risk

Borrowing authority creates strictly greater security risk than payment authority.

An agent compromised through prompt injection that possesses only payment authority can lose what it currently holds. An agent with borrowing authority can create liabilities exceeding its current assets, potentially pledging collateral, accepting unfavorable terms, or transferring borrowed funds to an attacker. The attack surface is strictly larger.25

No observed agent security architecture specifically addresses the separation of borrowing authority from payment authority. Current agent authorization frameworks (A2A, MCP, AP2) define spending limits but not debt limits. The IETF Agent Authorization Envelope draft defines budget constraints but does not distinguish credit from spending.

The FSB's warning that autonomous agent financial risks "can materialise at great speed" applies with greater force to credit. Regulators are moving toward tighter controls on autonomous financial activity, making delegated borrowing authority less likely in regulated environments.26

Default and Recourse

Credit becomes economically meaningful when non-repayment is possible. What happens when an agent defaults?

Potential default events include: insufficient balance, agent shutdown, identity abandonment, revenue collapse, principal intervention, and loss of credentials. An agent can be terminated by its principal, its infrastructure can fail, its revenue source can disappear, or it can simply be shut down and restarted under a new identity.

Potential recourse for a lender: seize collateral (in overcollateralized lending), sweep revenue (requires control of the payment channel), damage reputation (requires persistent identity), block marketplace access (platform-dependent), or pursue the legal principal.

The critical test: if every meaningful remedy ultimately targets a human, company, or collateral rather than the agent itself, then the agent is not the true borrower regardless of which software initiated the transaction. In overcollateralized lending, recourse is against collateral. In principal-backed credit, recourse is against the principal. In platform credit, recourse is through revenue sweep. In none of these cases is the agent itself the target of enforcement.

Competing Hypotheses

The evidence supports multiple viable interpretations:

Implications for the Four Scenarios

What Would Change Our Mind

We would revise the assessment that no lender underwrites agent economic capacity if a production lending system demonstrated credit decisions that depend materially on the autonomous agent's expected future economic activity, not on collateral value or principal creditworthiness.

We would reconsider the principal-backed credit hypothesis if a jurisdiction enacted legislation recognizing AI agents as independent debtors capable of bearing financial liability.

We would reassess the identity obstacle if a widely adopted identity system made wallet or agent identity costly to abandon, demonstrating Sybil resistance in a lending context.

We would reconsider the null hypothesis if identified autonomous agent commerce exceeded $1 million/month, creating sufficient scale for credit markets to form.

We would reconsider the platform-credit hypothesis if an agent marketplace or platform launched a revenue-based financing product for agent operators or agents themselves.

Known and Unknown

Known

• Software can technically initiate borrowing transactions and manage collateral
• Overcollateralized DeFi lending is permissionlessly accessible to any wallet
• Flash loans are not credit (no intertemporal risk, atomic repayment)
• Undercollateralized DeFi lending to known institutions has largely failed
• Legal liability routes to the deploying human/corporation in all examined jurisdictions
• On-chain credit scoring exists (Credora has facilitated $1.5B) but has no demonstrated predictive power for agent borrowers
• Wallet identity is pseudonymous and cheaply disposable; ERC-8004 provides agent identity infrastructure but does not make it costly to abandon
• Median agent lifespan is 2.48 minutes; 87% of agents die within 72 hours
• The only attempted autonomous agent loan (Bank of Bots / Freysa) resulted in total loss within five months
• Agent payment infrastructure is designed for immediate settlement, not credit
• Autonomous agent commerce is $5K-$11K/month, insufficient to generate credit demand
• Regulators are tightening controls on autonomous financial activity

Unknown

• Whether persistent Level 3 agents will emerge with sufficient economic activity to generate credit demand
• Whether agent revenue can be verified reliably enough to underwrite
• Whether on-chain credit scores can be validated against actual repayment data
• Whether platforms will extend Shopify Capital-style financing to agent commerce
• Whether trade credit will emerge between agents once commerce reaches sufficient scale
• Whether any jurisdiction will create a legal framework for AI agent debt capacity (Singapore's MAS SAFR may be a leading indicator)
• Whether delegated underwriting mechanisms can enable pseudonymous agent borrowing with operator risk absorption
• Whether agent-managed DAO LLCs can function as borrowing vehicles without human members
• Whether identity systems (SBTs, DIDs) will make machine identity costly to abandon
• Whether machine-readable financial state reduces underwriting cost enough to enable micro-credit
• The volume of agent economic activity on conventional rails (unmeasured)

Research Dependencies

This investigation builds on findings from Research 001 (economic agency framework, $5K-$11K autonomous commerce), Research 002 (wallet architecture), Research 003 (identity vs. authority), Research 004 (ownership), Research 005 (autonomous corporations, legal personality), Research 006 (payment rails), Research 007 (treasury, spending authority), and Research 008 (agent-to-agent commerce, trade relationships).

It hands off to Research 010 (Machine Capital Markets): if agent credit produces tradable debt claims, secondary markets, securitization, and broader capital-market implications become relevant.